Rishabh Shah.

1 problem solver, drawn to scale

Rishabh
Shah.

• 01Rebuilds physician and consult state by replaying an event stream server-side, with an injected clock, so every decision is deterministic and explainable.
• 02Detect, decide, act: flags breaching time standards, demand surges, and coverage gaps; picks the cheapest safe fix (reassign, escalate, wait, or a short split shift) and names the approver.
• 03Act loop runs propose → clinical-lead approval → send → accept, decline, or 10-minute timeout → recompute coverage, with every step in an append-only PostgreSQL audit log.
• 04FastAPI backend with an OpenAPI contract that generates the React dashboard's types; a Flutter app gives the clinical lead and physicians their own role-based views.

• 01Turns a free-text description of the day into an ordered itinerary, timed around travel between stops and each place's opening hours.
• 02Estimates crowds from MTA subway ridership near each stop, and schedules every stop for its quiet hour.
• 03Shows the plan on a 3D city from Google Maps Platform, with places and streets from OpenStreetMap, weather from Open-Meteo, and photos from Wikimedia Commons.

My role: Sole developer. Engineered a local-first Python developer tool spanning a CLI, MCP stdio server, SDK facade, SQLite graph index, and VS Code wrapper to generate compact repository context for coding agents.

• 01Tree-sitter indexing pipeline for Python, JavaScript, and TypeScript extracts files, symbols, imports, calls, inheritance, tests, and source references into a transactional SQLite graph.
• 02Adaptive retrieval combines FTS/BM25, graph proximity, token budgets, and strict schemas to return task-scoped context.
• 03Surfaces compact context through 9 CLI commands and 6 strict-schema MCP tools, shipped as a PyPI package.
• 04Crash-safe re-indexing: SHA-256 change detection re-parses only what changed, then swaps the index atomically so agents never read a half-built graph.
• 05Every release is gated on PR and nightly evaluation suites across 12 open-source repos (12,245 files), including Django, pandas, and Transformers.

• 01Questions come from a curated corpus through RAG over pgvector, with idempotent ingestion and a retrieval verification script, so interviews stay grounded instead of improvised.
• 02Gemini returns strict JSON Schema output with exact-quote citations; the Go server computes weighted verdicts across structure, clarity, depth, and ownership, and a cleanup job removes invalid scores.
• 03Recordings upload directly to S3 as 5 MiB multipart chunks with retry and part reconciliation, capping in-browser memory at 32 MiB per recording.
• 04Postgres-backed rate limits, CSRF checks, and per-user presigned S3 keys protect 4 endpoint types.

My role: Co-author. Built the OCR step (Llama 4 Scout vision, a retry on the next sharpest frame, and regex checks for Indian plates) and helped fine-tune YOLO12-X, the best of 5 YOLO11/YOLO12 variants.

• 01Integrated Llama 4 Scout via Groq for license plate OCR, regex-based plate validation, vehicle registration verification, Jinja2/pdfkit document generation, AWS S3 storage, and Twilio messaging.
• 02Violation checks fire only when a tracked vehicle crosses a virtual line instead of on every frame of 60 fps video, cutting wasted compute and duplicate tickets. ByteTrack keeps per-vehicle state across frames.
• 03YOLOv12x fine-tuned on a custom traffic dataset covering red light jumping, speeding, helmet-less riding, missing insurance, and pollution non-compliance.

My role: First author. Designed and implemented the full pipeline: WGAN-GP architecture, 3000-epoch training, synthetic-data filtering, and downstream prediction evaluation on unseen data.

• 01Conditional Wasserstein GAN with gradient penalty plus mean-variance matching, not a vanilla GAN. It stabilizes 3,000 epochs of adversarial training on high-variance series where vanilla GANs collapse.
• 02Causal TCN layers feed stacked LSTMs, so the generator never sees the future while still learning long-range structure across 6 market features modeled as conditional sequences.
• 03Generator concatenates past market windows with noise, then runs multi-head attention, temporal convolution blocks, two LSTM layers, and a GRU before a dense head. The critic is a Conv1D plus LeakyReLU stack trained 5 times per generator step with AdamW.
• 04Trained on AAPL daily data (Open, High, Low, Close, Adjusted Close, Volume) preprocessed with log returns and a hybrid standard plus min-max scaling scheme. Synthetic sequences are validated with Pearson, Spearman, and R-squared before being mixed into training sets.

My role: End-to-end build: Django backend with authentication via django-allauth, Groq API integration for LLM inference, Raspberry Pi 3B+ device control, Azure IoT Hub messaging, and the wake-word speech pipeline.

• 01Four modules: deterministic keyword tasks, multi-threaded math and reminder handling, LLM query processing, and home automation. The wake word JARVIS triggers the speech pipeline after noise filtering.
• 02LLM queries are forwarded with guard prompts to Llama 3.1 8B hosted on Groq; prompt engineering filters harmful context and prevents exposure of private data before responses reach the user.
• 03Home automation commands flow through Azure IoT Hub to a Raspberry Pi 3B+ that toggles appliances via GPIO pins, with device state mirrored in a database and voice commands never cached on device.

My role: Built the full pipeline: feature extraction (MFCC plus spectral features via Librosa), Bi-LSTM architecture in TensorFlow, training loop, post-training quantization, and a Streamlit demo app.

• 01Librosa extracts MFCC coefficients plus spectral centroid and rolloff. Features are normalized and fed into a 2-layer Bi-LSTM with dropout between layers.
• 02Post-training quantization shrinks the model by 15% and speeds inference by 22% at negligible accuracy cost. Deploy-time is materially better.
• 03Streamlit demo app uploads audio, extracts features on the fly, runs inference, and shows per-genre probability plus a confidence spark chart.